Gimmenoto/Warinza008

♟️ Lichess Elite Chess Tensor Dataset (2013 – May 2020)

Overview

This dataset contains high-dimensional chess board representations extracted exclusively from the Lichess Elite Database, covering games from January 2013 to May 2020.

Each chess position is encoded as a structured 125 × 8 × 8 float tensor, designed for deep learning, policy learning, and reinforcement learning research.

This dataset was prepared strictly for educational and academic research purposes only, under the Department of Computer Science,
King Mongkut’s University of Technology Thonburi (KMUTT), Thailand.

---

Data Source

Original dataset:

• Lichess Elite Database
  https://database.nikonoel.fr/

Time range used:

• January 2013 – May 2020

No data outside this range is included.

This dataset is not affiliated with Lichess. All rights remain with the original data provider.

---

Dataset Structure

Split

Split	Description
train	All extracted board positions

---

Features

Column	Type	Description
board	list<list<list>>	125-channel 8×8 tensor
win	bool	True if game result is 1-0
draw	bool	True if game result is 1/2-1/2
lose	bool	True if game result is 0-1
piece_to_move	int64	Encoded move target (from_square × 64 + to_square)

---

Board Tensor Specification (125 Channels)

The tensor is composed of structured feature groups described below.

---

1. Piece Planes (14 channels)

Binary bitboards (1.0 where piece exists, else 0.0).

White pieces:

• Pawn
• Knight
• Bishop
• Rook
• Queen
• King

Black pieces:

• Pawn
• Knight
• Bishop
• Rook
• Queen
• King

Additional duplicated bishop planes:

• White bishops
• Black bishops

Total: 14 channels

---

2. Castling Rights (4 channels)

Each channel is a full 8×8 plane filled with 1.0 if the right exists.

• White kingside
• White queenside
• Black kingside
• Black queenside

---

3. En Passant (2 channels)

• En passant target square mask
• En passant capture pawn square

---

4. Game State Scalars (4 channels)

Each is broadcasted over the entire 8×8 plane.

• Side to move (1.0 = White, 0.0 = Black)
• Halfmove clock normalized (halfmove_clock / 100)
• Repetition ≥ 2
• Repetition ≥ 3

---

5. Tactical Information (6 channels)

• Checkers mask
• Squares attacked by White
• Squares attacked by Black
• White pinned pieces
• Black pinned pieces
• Threatened pieces (current side pieces under attack)

---

6. King Mobility (2 channels)

• White king legal moves
• Black king legal moves

---

7. Pawn Structure (5 channels)

• White passed pawns
• Black passed pawns
• Isolated pawns
• Doubled pawns
• Structural pressure proxy

---

8. History Planes (84 channels)

Last 6 board positions × 14 piece channels each.

Order:

• Most recent position first

Each historical board encodes:

• 6 white piece planes
• 6 black piece planes
• 2 bishop duplicate planes

---

9. Move Masks and Evaluation (4 channels)

• Legal move destination squares
• Capture squares
• Promotion squares
• Mobility scalar (legal_moves / 218)

---

10. Material Balance (1 channel)

Normalized material difference:

Piece values:

• Pawn = 1
• Knight = 3
• Bishop = 3
• Rook = 5
• Queen = 9

---

Channel Summary

Category	Channels
Pieces	14
Castling	4
En passant	2
Game state	4
Tactical info	6
King mobility	2
Pawn structure	5
History	84
Move masks + mobility	4
Material balance	1
TOTAL	125

---

Intended Use

This dataset is suitable for:

• Policy prediction (move selection)
• WDL classification (win/draw/lose)
• Value networks
• AlphaZero-style supervised pretraining
• Position evaluation modeling
• Tactical learning

---

Example Usage (PyTorch)

from datasets import load_dataset
import torch

dataset = load_dataset("your-username/lichess-elite-125ch")

sample = dataset["train"][0]

board_tensor = torch.tensor(sample["board"])  # shape [125, 8, 8]
move_target = sample["piece_to_move"]